feat(langfuse): propagate observation metadata for evals (#14634)

# Pull Request Template

## Description

We need to pass on trace level attributes down to the spans inside them
like tool calls, observations, etc.
This way, we can filter observations based on trace level attributes.


## Type of change

- [x] Bug fix (non-breaking change which fixes an issue)

## How Has This Been Tested?

Please describe the tests that you ran to verify your changes. Provide
instructions so we can reproduce. Please also list any relevant details
for your test configuration.

Attributes added to observation metadata for easy filtering
<img width="1327" height="708" alt="image"
src="https://github.com/user-attachments/assets/8f1d1bf8-cde4-481d-a2c2-7920ad2fc52e"
/>

added a `generation_stage` to differentiate llm_calls that call tools vs
those that generate a `final_response`
<img width="1806" height="968" alt="CleanShot 2026-06-03 at 15 11 09@2x"
src="https://github.com/user-attachments/assets/db1fa8e0-7f2d-404b-a719-27a16d400442"
/>


propagated attributes to tool calls for future use
<img width="903" height="517" alt="image"
src="https://github.com/user-attachments/assets/edc61ce8-93db-465c-a66e-043138e2dc15"
/>



## Checklist:

- [x] My code follows the style guidelines of this project
- [x] I have performed a self-review of my code
- [x] I have commented on my code, particularly in hard-to-understand
areas
- [ ] I have made corresponding changes to the documentation
- [x] My changes generate no new warnings
- [x] I have added tests that prove my fix is effective or that my
feature works
- [x] New and existing unit tests pass locally with my changes
- [x] Any dependent changes have been merged and published in downstream
modules

spec/enterprise/services/captain/llm/assistant_chat_service_spec.rb

@@ -39,6 +39,30 @@ RSpec.describe Captain::Llm::AssistantChatService do
       allow(mock_chat).to receive(:ask).and_return(mock_response)
       service.generate_response(message_history: [{ role: 'user', content: 'Hello' }])
     end
+
+    it 'marks final response generations for observation-level evaluators' do
+      service = described_class.new(assistant: assistant, conversation: conversation)
+      message = instance_double(RubyLLM::Message, content: 'Final answer', input_tokens: 10, output_tokens: 20, tool_calls: {})
+
+      attributes = service.send(:generation_attributes, mock_chat, message)
+
+      expect(attributes['langfuse.observation.metadata.generation_stage']).to eq('final_response')
+    end
+
+    it 'marks tool call generations separately from final responses' do
+      service = described_class.new(assistant: assistant, conversation: conversation)
+      message = instance_double(
+        RubyLLM::Message,
+        content: '',
+        input_tokens: 10,
+        output_tokens: 20,
+        tool_calls: { 'call_1' => instance_double(RubyLLM::ToolCall) }
+      )
+
+      attributes = service.send(:generation_attributes, mock_chat, message)
+
+      expect(attributes['langfuse.observation.metadata.generation_stage']).to eq('tool_call')
+    end
   end
 
   describe 'image analysis' do

spec/lib/integrations/llm_instrumentation_spec.rb

@@ -144,7 +144,10 @@ RSpec.describe Integrations::LlmInstrumentation do
 
         expect(mock_span).to have_received(:set_attribute).with('langfuse.user.id', '123')
         expect(mock_span).to have_received(:set_attribute).with('langfuse.session.id', '123_456')
-        expect(mock_span).to have_received(:set_attribute).with('langfuse.trace.tags', '["reply_suggestion"]')
+        expect(mock_span).to have_received(:set_attribute).with('langfuse.trace.tags', ['reply_suggestion'])
+        expect(mock_span).to have_received(:set_attribute).with('langfuse.observation.metadata.user_id', '123')
+        expect(mock_span).to have_received(:set_attribute).with('langfuse.observation.metadata.session_id', '123_456')
+        expect(mock_span).to have_received(:set_attribute).with('langfuse.observation.metadata.feature_name', 'reply_suggestion')
       end
 
       it 'sets completion message attributes when result contains message' do
@@ -253,6 +256,76 @@ RSpec.describe Integrations::LlmInstrumentation do
           expect(mock_span).to have_received(:set_attribute).with('langfuse.observation.output', result_data.to_json)
         end
 
+        it 'propagates trace attributes as observation metadata to child tool spans' do
+          root_span = instance_double(OpenTelemetry::Trace::Span)
+          tool_span = instance_double(OpenTelemetry::Trace::Span)
+          tool_instance = test_class.new
+          allow(root_span).to receive(:set_attribute)
+          allow(tool_span).to receive(:set_attribute)
+          allow(instance).to receive(:tracer).and_return(mock_tracer)
+          allow(tool_instance).to receive(:tracer).and_return(mock_tracer)
+          allow(mock_tracer).to receive(:in_span).with('llm.test').and_yield(root_span)
+          allow(mock_tracer).to receive(:in_span).with('tool.search').and_yield(tool_span)
+
+          instance.instrument_agent_session(params) do
+            tool_instance.instrument_tool_call('search', { query: 'test' }) { 'tool result' }
+          end
+
+          expect(tool_span).to have_received(:set_attribute).with('langfuse.observation.metadata.user_id', '123')
+          expect(tool_span).to have_received(:set_attribute).with('langfuse.observation.metadata.session_id', '123_456')
+          expect(tool_span).to have_received(:set_attribute).with('langfuse.observation.metadata.feature_name', 'reply_suggestion')
+        end
+
+        it 'keeps inherited session metadata for nested service spans with their own feature tag' do
+          root_span = instance_double(OpenTelemetry::Trace::Span)
+          nested_span = instance_double(OpenTelemetry::Trace::Span)
+          nested_instance = test_class.new
+          nested_params = params.merge(span_name: 'llm.translate_query', conversation_id: nil, feature_name: 'translate_query')
+          allow(root_span).to receive(:set_attribute)
+          allow(nested_span).to receive(:set_attribute)
+          allow(instance).to receive(:tracer).and_return(mock_tracer)
+          allow(nested_instance).to receive(:tracer).and_return(mock_tracer)
+          allow(mock_tracer).to receive(:in_span).with('llm.test').and_yield(root_span)
+          allow(mock_tracer).to receive(:in_span).with('llm.translate_query').and_yield(nested_span)
+
+          instance.instrument_agent_session(params) do
+            nested_instance.instrument_llm_call(nested_params) { 'translated query' }
+          end
+
+          expect(nested_span).to have_received(:set_attribute).with('langfuse.session.id', '123_456')
+          expect(nested_span).to have_received(:set_attribute).with('langfuse.trace.tags', ['translate_query'])
+          expect(nested_span).to have_received(:set_attribute).with('langfuse.observation.metadata.session_id', '123_456')
+          expect(nested_span).to have_received(:set_attribute).with('langfuse.observation.metadata.feature_name', 'translate_query')
+        end
+
+        it 'propagates session metadata to nested embedding spans' do
+          root_span = instance_double(OpenTelemetry::Trace::Span)
+          embedding_span = instance_double(OpenTelemetry::Trace::Span)
+          embedding_instance = test_class.new
+          embedding_params = {
+            span_name: 'llm.captain.embedding',
+            account_id: 123,
+            feature_name: 'embedding',
+            model: 'text-embedding-3-small',
+            input: 'search result'
+          }
+          allow(root_span).to receive(:set_attribute)
+          allow(embedding_span).to receive(:set_attribute)
+          allow(instance).to receive(:tracer).and_return(mock_tracer)
+          allow(embedding_instance).to receive(:tracer).and_return(mock_tracer)
+          allow(mock_tracer).to receive(:in_span).with('llm.test').and_yield(root_span)
+          allow(mock_tracer).to receive(:in_span).with('llm.captain.embedding').and_yield(embedding_span)
+
+          instance.instrument_agent_session(params) do
+            embedding_instance.instrument_embedding_call(embedding_params) { [0.1, 0.2, 0.3] }
+          end
+
+          expect(embedding_span).to have_received(:set_attribute).with('langfuse.session.id', '123_456')
+          expect(embedding_span).to have_received(:set_attribute).with('langfuse.trace.tags', ['embedding'])
+          expect(embedding_span).to have_received(:set_attribute).with('langfuse.observation.metadata.session_id', '123_456')
+          expect(embedding_span).to have_received(:set_attribute).with('langfuse.observation.metadata.feature_name', 'embedding')
+        end
+
         # Regression test for Langfuse double-counting bug.
         # Setting gen_ai.request.model on parent spans causes Langfuse to classify them as
         # GENERATIONs instead of SPANs, resulting in cost being counted multiple times